You would not use spoon 🥄 as a fork 🍴 or a knife. Why would data scientists try to use forecasting to solve anomaly detection problems. This is so wrong. Anomaly detection is an own domain that has its own methods. One of the pitfalls of #datascientists is trying to treat #anomalydetection as a forecasting problem and use forecasting models to try to detect anomalies using forecasting methods. This is wrong, an example below is a simple illustration of why such methods don’t work, in a simple time series two artificial anomalies were injected dropping signal value to zero. For the first anomaly, the value of zero is still within “prediction interval” (which is most likely to be wrong as well unless conformal prediction has been used but this is not the subject of this post). Forecasting based model then sees the value of zero as within prediction interval and clearly won’t label 🏷️ this anomaly as anomaly that it clearly is. Don’t treat anomaly detection problems as forecasting problems, learn about anomaly detection as its own domain and what works and what does not work there. Follow Eamonn Keogh works as he is #1 Professor globally in the field of anomaly detection. Look at libraries like STUMPY created by @seanmylaw to deal with anomaly detection problems. Learn about how #conformalprediction has successfully tackled anomaly detection (conformal prediction powers Microsoft Azure anomaly detection module) and other methods that work. You would not use spoon 🥄 as a fork 🍴 or a knife. @seanmylaw has got you covered, the best series of articles on matrix profile (research by Eamonn Keogh) explaining the matrix profile methods in plain English medium.com/@seanmylaw #anomalydetection #machinelearning #timeseries